Sources for ICDM2018: EPAB: Early Pattern Aware Bayesian Model for Social Content Popularity Prediction
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
paper
pic
src
ICDM_cover.png
README.md

README.md

ICDM2018-EPAB

Conference: IEEE International Conference on Data Mining (Singapore) - ICDM2018 (accepted)
Author: Qitian Wu, Chaoqi Yang, Xiaofeng Gao, Peng He, Guihai Chen
Title: EPAB: Early Pattern Aware Bayesian Model for Social Content Popularity Prediction



Brief Model

(1). The early pattern of each cascade is represent as a vector:


cluster the early pattern


(2). We introduce three hidden variable to capture describe the state of each cascade.

  • Influence(h1): how many people have been influenced by this tweet.
  • Attractiveness(h2): how many people tend to click and repost this tweet.
  • Potentiality(h3): how many people will be exposed to this tweet.
  • optimize the loss function between ground truth and predicted value
  • get alpha, beta, gamma for each pattern
  • and get h1, h2, h3 for each cascade.

(3). two-layer Bayesian network to model observable feature X, hidden variable H, and final state Y.

  • refer to the paper for detailed deduction
  • bayesian rule

(4). Solve the loss function, get the model.

  • to solve this loss function, we can compute the loss of theta1, theta3, and theta2 seperately
  • use stochastic gradient descent and hill-climbing


Edited on Sep. 10th, 2018